Two k-winners-take-all networks with discontinuous activation functions

Area of Science:

• Computational neuroscience
• Artificial neural networks

Background:

• The k-winners-take-all (k-WTA) operation is crucial for feature selection and pattern recognition in neural networks.
• Existing k-WTA implementations often face challenges with discontinuous activation functions and real-time processing.

Purpose of the Study:

• To develop novel k-winners-take-all (k-WTA) networks capable of handling discontinuous activation functions.
• To design networks that can perform the k-WTA operation efficiently and in real time.

Main Methods:

• Formulating the k-WTA operation as equivalent linear and quadratic programming problems.
• Designing two distinct k-WTA network architectures based on these programming formulations.
• Utilizing discontinuous activation functions within the network designs.

Main Results:

• Theoretical guarantees for real-time k-WTA operation.
• Demonstrated effectiveness and performance through simulation results.
• Successful implementation of k-WTA using discontinuous activation functions.

Conclusions:

• The proposed k-WTA networks offer a viable solution for real-time computation with discontinuous activation functions.
• The linear and quadratic programming approaches provide a robust foundation for designing efficient k-WTA networks.
• Simulation results validate the practical applicability and performance of the developed networks.